Orthopaedics Example

* Lossless compression to 4.9% of the original image.
* Decompression to lower resolution, or region of interest blocks at full resolution, in near real time.
* Regions of interest can be reconstructed in near real time from block tiling.

The following image is originally of dimensions 1760 x 2140, in DICOM 16-bit format, and weighing in at 7.5 MB. In the current implementation we converted to, and subsequently used, 32-bit FITS, giving an image size of 15 MB, to facilitate floating point computation, and times quoted here are all based on this bloated image size. Timings were carried out on a 360 MHz Sun Sparcstation (which is a reasonable but by no means very powerful workstation in today's marketplace).

The image shows a fractured tibia, with an intramedullary nail (the long, dark artificial structure) inserted in the bone. On the right are a number of recepticles containing crushed bone to provide calibration standards to enable bone mineral density to be estimated from image graylevel measurements. During the healing period, bone tissue is regenerated at the fracture site. In the approximate 12 weeks of healing, x-ray images are used to check the patient's progress every few weeks.

The image shown is at a resolution scale which is twice times twice reduced in both dimensions. Lossless compression took 6 seconds on this 15 MB image, yielded an image of size 4.9% of the input image, and allows decompression in well under 0.1 second. (In all of our timing tests, the system time was 0.0 seconds.) The decompression timing tests apply to (i) the reduced resolution scale shown here, or (ii) decompression at full resolution of a 128 x 128 block.

The fracture area is the most important immediate aspect of such an image. Using 128 x 128 blockwise compression allowed us to build up a set of 9 of these blocks at full resolution, in a 3 x 3 configuration around the "point of insult". The following, of overall dimensions 384 x 384, shows this set of 9 blocks, with histogram equalization to show up fine detail.

Further development possibilities: (i) Automatically find fracture location. (ii) Generate a reduced-resolution version of the image. (iii) Compress, blockwise, the image. (iii) On demand, provide a region around the fracture location, based on a grid of blocks as exemplified above. To be supported are format conversions and display operations such as histogram equalization.

Support for a library of tibia fracture images, on a web server or as a stand-alone version with a customized user interface for use on a portable computer, can be easily envisaged.